AI Framework Synthesizes Tool-Use Data for LLMs
Analysis
This paper addresses a significant challenge in enabling Large Language Models (LLMs) to effectively use external tools. The core contribution is a fully autonomous framework, InfTool, that generates high-quality training data for LLMs without human intervention. This is a crucial step towards building more capable and autonomous AI agents, as it overcomes limitations of existing approaches that rely on expensive human annotation and struggle with generalization. The results on the Berkeley Function-Calling Leaderboard (BFCL) are impressive, demonstrating substantial performance improvements and surpassing larger models, highlighting the effectiveness of the proposed method.
Key Takeaways
- •InfTool is a fully autonomous framework for generating tool-use data for LLMs.
- •It uses a multi-agent role-playing approach to create diverse and verified trajectories.
- •The framework establishes a closed loop, iteratively improving the model and data quality.
- •Achieves significant performance gains on the Berkeley Function-Calling Leaderboard (BFCL).
- •Demonstrates the potential of synthetic data for training LLMs in tool use.
“InfTool transforms a base 32B model from 19.8% to 70.9% accuracy (+258%), surpassing models 10x larger and rivaling Claude-Opus, and entirely from synthetic data without human annotation.”